Grant: Discovery Grants Program - Individual

Hydropower is a crucial industry in Canada, providing over 60% of the country's electricity needs. As we transition to a greener future, it is imperative that we maximize the efficiency of all hydropower systems as much as possible. Hydropower reservoir management requires multiple steps that must be chained together to provide the best turbine operation schedule. First, weather forecasts are generated by weather forecasting centers. These forecasts are sometimes biased or do not represent weather patterns reliably. These forecasts are then used in hydrological models that transform the weather into flow forecasts. These models are imperfect simplifications of watersheds and thus introduce errors and uncertainty in the process. Finally, these flow forecasts are used as possible sets of inflows by reservoir optimization software, that must be calibrated and tuned to specific hydropower systems. Each step thus adds a layer of complexity that can impact final hydropower output. To make things worse, human intervention is often required to adjust flow forecasts, correct hydrological model drifts, and try to reduce cumulative errors. However, this is imprecise and depends on the engineer that is performing the adjustment. Recently, machine learning models have been developed to tackle different aspects of this chain. This paradigm shift has shown that these models can outperform our classical models in some circumstances. They can be trained to make optimal decisions from streamflow forecasts, and they can forecast flows from weather forecasts. Furthermore, they do not require human intervention once trained. This research program aims to automate the hydropower reservoir management process by developing improvements to machine learning models in each sub-process (hydrological forecasting and decision making), and finally coupling them into a single model that will allow providing water management decisions directly from weather forecasts. This will have the advantage of including all biases internally from the training session, leading to synergetic gains. Furthermore, machine learning models can make use of much more data than typical hydrological models. To this end, this research program will make use of large-scale databases of weather observations, forecasts, and streamflow to train large machine learning models over thousands of catchments. This will allow using these models for flow forecasting and decision-making on any catchment and hydropower system. This research is important for the hydropower industry as well as the public safety institutions from improvements in flood forecasting reliability. It will lead to economic benefits for the industry, reduced dependence on fossil fuels and greenhouse gas emissions and safer dam management, along with the training of 8 highly qualified personnel. Finally, it will also lead to a new paradigm in how hydropower reservoirs are operated, which will open the door to numerous future research avenues.